sensors Article Improved Pedestrian Dead Reckoning Based on a Robust Adaptive Kalman Filter for Indoor Inertial Location System Qigao Fan 1,†, Hai Zhang 1,*,†, Peng Pan 2,†, Xiangpeng Zhuang 1, Jie Jia 1, Pengsong Zhang 1, Zhengqing Zhao 1, Gaowen Zhu 1 and Yuanyuan Tang 1 1 Internet of Things Engineering, Jiangnan University, Wuxi 214000, China; [email protected] (Q.F.); [email protected] (X.Z.); [email protected] (J.J.); [email protected] (P.Z.); [email protected] (Z.Z.); [email protected] (G.Z.); [email protected] (Y.T.) 2 Department of Mechanical Engineering, McGill University, Montreal, QC H3A 0G4, Canada; [email protected] * Correspondence: [email protected]; Tel.: +86-188-0059-0310 † These authors contributed equally to this work. Received: 23 November 2018; Accepted: 8 January 2019; Published: 12 January 2019 Abstract: Pedestrian dead reckoning (PDR) systems based on a microelectromechanical-inertial measurement unit (MEMS-IMU) providing advantages of full autonomy and strong anti-jamming performance are becoming a feasible choice for pedestrian indoor positioning. In order to realize the accurate positioning of pedestrians in a closed environment, an improved pedestrian dead reckoning algorithm, mainly including improved step estimation and heading estimation, is proposed in this paper. Firstly, the original signal is preprocessed using the wavelet denoising algorithm. Then, the multi-threshold method is proposed to ameliorate the step estimation algorithm. For heading estimation suffering from accumulated error and outliers, robust adaptive Kalman filter (RAKF) algorithm is proposed in this paper, and combined with complementary filter to improve positioning accuracy. Finally, an experimental platform with inertial sensors as the core is constructed. Experimental results show that positioning error is less than 2.5% of the total distance, which is ideal for accurate positioning of pedestrians in enclosed environment. Keywords: indoor inertial positioning; MEMS-IMU; improved pedestrian dead reckoning; robust adaptive Kalman filter 1. Introduction Indoor positioning is a technology for positioning in an indoor environment, which shows great prospects in industries such as emergency rescue, logistics, wireless games and shopping malls [1–3]. The Global Navigation Satellite System, such as the Global Positioning System (GPS), BeiDou Navigation Satellite System (BDS), and Global Navigation Satellite System (GLONASS), can provide absolute accurate positioning, but its performance in some enclosed environments, such as tunnels and mines, is not satisfied; as the Global Navigation Satellite System may suffer heavily from satellite signal blockage [4]. Therefore, how to obtain accurate position information of indoor objects has attracted wide attention. At present, indoor positioning methods have been proposed, including Ultra Wideband (UWB) [5], ZigBee [6], Wi-Fi [7], Bluetooth [8], Radio Frequency Identification (RFID) [9], and visual positioning [10]. However, performances of these positioning methods are easily affected by external environment. UWB positioning is susceptible to non-line-of-sight interference such as occlusion, external collision and strong magnetic field interference [11]. Wi-Fi positioning is susceptible Sensors 2019, 19, 294; doi:10.3390/s19020294 www.mdpi.com/journal/sensors Sensors 2019, 19, 294 2 of 21 to interference from other signals and consumes a lot of power. In addition, ZigBee and Bluetooth positioning are not suitable for complex environments, in which the performance is not stable. RFID positioning does not have communication capabilities, and the anti-interference ability is poor. Visual positioning has poor real-time performance and is limited by light conditions, which cannot work in dark environments [12]. With the continual miniaturization of Micro-Electro-Mechanical Systems (MEMS) in recent years, the MEMS-IMU has the advantages of simple structure and portability. At the same time, inertial measurement is completely autonomous, suitable for all weather, free from external environment interference, and has no signal loss, so MEMS-IMU-based inertial navigation technology is very suitable for indoor positioning [13,14]. Pedestrian dead reckoning (PDR) is an effective technique for indoor positioning of pedestrians [15–17]. By measuring the movement information of pedestrians in real time, the pedestrian’s step, step length, and heading angle are identified to calculate the position of the pedestrian. The indoor positioning based on MEMS-IMU is widely used because of its advantages such as portability, simple structure, and strong anti-interference performance. However, errors of the Inertial positioning system accumulated over time, which greatly affects the positioning accuracy [18]. Domestic and foreign experts have proposed combined positioning system to suppress cumulative errors through auxiliary positioning techniques, such as INS/GNSS [19], INS/RFID [20], and INS/UWB positioning [21]. The combined positioning technology can improve the positioning accuracy compared with the single inertial positioning technology, but its external requirements are more demanding, which greatly limits the scope of use of the combined positioning system. In scenes with harsh external environments, such as the disaster scene of mines or tunnels, the combined positioning system is difficult to build and wastes valuable rescue time. The single inertial positioning technology is extremely suitable because of its characteristics, so the study of single inertial positioning system has high practical value. In the framework of pedestrian dead reckoning, the step estimation is very important. The key to correctly estimate step frequency is to avoid acceleration warpage and sensor noise, which may lead to incorrect step estimation. Noises usually exit in the signal obtained by MEMS-IMU, which result in the poor positioning result. To improve accuracy of positioning, it is necessary to preprocess the signal. Low pass filter is typically used to smooth the data, but it is difficult to determine the appropriate cutoff frequency [22]. The authors in [23] analyze error noise term by Allan variance method, and optimized Kalman filter is designed to denoise the signal according to noise correlation, but the amount of calculation is large. At present, the step estimation algorithm usually includes Peak detection [24–26], Zero crossings [27] and fast Fourier transform [28], where the signals of accelerometer are generally used. In these algorithms, the placement of the IMU defaults to one of the accelerometer axes being always vertical [29], which greatly affects the shape of the accelerometer signal and the effectiveness of the algorithm. Step length estimation plays an important role in the pedestrian dead reckoning. For different individuals, the step length of different individuals varies. Even for the same person, step length will change due to different states of motion. Usually, there are two kinds of models for estimating the step length: static model and dynamic model. It is assumed that in the static model, the step length is a fixed value, which is obviously inaccurate. In comparison, the dynamic model considers that any effective stride has a different step length, which is affected by the different individuals or different states of motion. The step dynamic model includes step-frequency linear model [30], nonlinear model [31], and adaptive learning step length model. A nonlinear model is used in [32] to accommodate the difference of step length between different pedestrians, which highly stable and moderately complex compared to other dynamic step models. Heading estimation is the core of the pedestrian dead reckoning algorithm. Calculating the heading angle using the magnetometer has the advantages of simple structure, low power consumption, and no cumulative errors [33]. However, magnetometers are susceptible to external environmental disturbances, especially in mines or garages, where there are many ferromagnetic Sensors 2019, 19, 294 3 of 21 materials. There are serious cumulative errors when using the gyroscope alone, especially at corners. The longer the time is, the larger the cumulative error will be [34]. In order to improve the heading accuracy of the inertial measurement unit, fusion filtering algorithms are proposed, such as Kalman filtering [35] and complementary filtering [36]. The authors in [37] use an extended Kalman filter based on two-levelSensors 2019, 19 quaternions, x FOR PEER REVIEW to fuse signals of acceleration, angular velocity and magnetic to3 of minimize 22 thefiltering cumulative [35] and error, complementary but ignores the filtering change [36]. in The noise authors characteristics. in [37] use Therefore,an extende thed Kalman adaptive filter filtering algorithmbased on is proposed.two-level quaternions In the literature to fuse [38 signals], the fuzzyof acceleration, adaptive Kalmanangular velocity filter is usedand magnetic to fuse the to signals of gyroscopeminimize the and cumulative accelerometer, error, but but it isignores limited the by thechange precision in noise of thecharacteristics. fuzzy control, Therefore, and the calculationthe amountadaptive is large. filtering Sage–Husa algorithm adaptiveis proposed. filtering In the [ 39literature] can correct [38], the the fuzzy statistical adaptive characteristics Kalman filter of is process noiseused and to fuse measurement the